Back to the changelog

• Security

  The model only sees pseudonyms

  If the agent's AI provider has Data Shield enabled, the sensitive fields of items consulted in chat are replaced by tokens (`[REDACTED_…]`) before every call to the model — across the whole conversation, including tool results and previous turns. The real value never reaches the provider.

• Design

  Transparent for chat users

  Users keep seeing real data in the replies; a shield icon in the chat header signals that shielding is active.

• Feature

  The agent stays fully operational

  If the model uses a shielded value to search items or propose a start, the real value is restored at execution time — searches work and items are created with the correct data, without the model ever seeing them. The pseudonym↔value mapping stays stable for the whole session, even after reopening the chat.

• Security

  No extra configuration, MCP included

  It's enough that the agent's provider has Data Shield enabled (Settings → AI Providers, premium feature); with shielding off, sensitive fields stay redacted as before. The `flow_agent` MCP tool exposed to external assistants inherits the same protection. Also fixed a shield-token collision that, in edge cases, could overwrite a map entry on AI tasks with the two-level shield (deterministic + LLM).